Addressing positivity violations in causal effect estimation using Gaussian process priors.

Area of Science:

• Epidemiology
• Biostatistics
• Causal Inference

Background:

• Causal inference in observational studies depends on assumptions like positivity.
• Positivity requires treatment probabilities bounded away from 0 and 1 for all covariate combinations.
• Violations lead to extrapolation and uncertainty in causal effect estimates.

Purpose of the Study:

• To develop a Gaussian process model for estimating treatment effects.
• To address practical violations of the positivity assumption in causal inference.
• To incorporate greater uncertainty in estimates where covariate overlap is limited.

Main Methods:

• Constructed a Gaussian process model for treatment effect estimation.
• Focused on handling violations of the positivity assumption.
• Assessed performance using simulation studies for bias and efficiency.

Main Results:

• The Gaussian process model provides a cohesive approach to estimating treatment effects.
• The method generates increased uncertainty estimates in regions with poor covariate overlap.
• Simulations demonstrated the approach's performance regarding bias and efficiency.

Conclusions:

• The proposed Gaussian process model offers a robust method for causal inference with positivity violations.
• It enhances the reliability of treatment effect estimates by reflecting uncertainty.
• Applied to critically ill female patients, it examined the effect of right heart catheterization.